we should enable data page checksums

[jberkus]

If patroni is initdb'ing, we should enable data page checksums because:

• it's a good idea
• performance impact is not noticeable under most circumstances
• prior to 9.5, it's required for pg_rewind to be safe to use

[CyberDem0n]

I guess we should solve this problem more general. We can create special section in the config.yaml file which will contain all needed parameters and options for initdb

[jberkus]

Why?

If Patroni is doing initdb, I don't see why it should give the user an option to turn off checksums.

[a1exsh]

Do we have any recent benchmark measuring performance penalty of enabling
data checksums? Last time I saw one, the effect was not nearly negligible,
I believe.

On Tue, Nov 3, 2015, 00:21 Josh Berkus notifications@github.com wrote:

Why?

If Patroni is doing initdb, I don't see why it should give the user an
option to turn off checksums.

—
Reply to this email directly or view it on GitHub
#87 (comment).

[jberkus]

Hmmm, I need to publish. On workloads I've tested overhead has been < 5%. Compared to the safety of detecting corruption as soon as it happens, that's negligable. Also, I'm pretty sure that pg_rewind without checksums is unsafe, and you're building pg_rewind into Patroni.

[a1exsh]

Would be glad to see the benchmark. :-) pg_rewind requires checksums OR
logging of hint bits iirc.

On Tue, Nov 3, 2015, 19:38 Josh Berkus notifications@github.com wrote:

Hmmm, I need to publish. On workloads I've tested overhead has been < 5%.
Compared to the safety of detecting corruption as soon as it happens,
that's negligable. Also, I'm pretty sure that pg_rewind without checksums
is unsafe, and you're building pg_rewind into Patroni.

—
Reply to this email directly or view it on GitHub
#87 (comment).

[alexeyklyukin]

Patroni can be attached to an already running cluster, or maybe in the future used to initialise new replicas out of the master that is completely unaware of patroni. On both of those cases we cannot guarantee the checksums will be on. We might solve this by fetching the checksum value parameter out of an existing master, but I think it would be easier to just make initdb options configurable.

[jberkus]

We don't initdb in either of those cases.

[alexeyklyukin]

Oh, right. But we also won't be able to rely on the checksums being on for pg_rewind in those cases. I think we should make this option configurable and choose the default value depending on the results of the benchmarks.

[feikesteenbergen]

Implemented with custom options.